QNAP TVS-472XT performance issue

[Laseru]

Hi,
I've purchased lately a QNAP TVS-472XT-PT-4G which seems to be a good upgrade over my older TVS-682 but not as good as expected.

I connected my NAS using a 10 GB switch to my PC which is also using a 10GB connection and by Thunderbolt 40GB/s to my iMac.

I will not focus on the speeds I get for the mechanical drives but only on the M2 SSDs which are Samsung Evo 970 1TB.
In the beginning I tried the performance using just one M2 SSD and my speeds topped at 1300 MB/s write for the Thunderbolt and at 1000 MB/s write by using the PC over 10GB/s.
I have no clue why but the read speeds are much lower, I was expecting the read speeds to be higher than the write speeds.
In the beginning I taught that because the M2 is using only 2 PCI Express lanes and theoretical speed is 2GB/s by adding the second M2 in a RAID 0 configuration I will increase the speeds.
But that is not true, I see zero performance gains when using two M2 SSDs and that I find weird, it seems like the speeds are almost identical.

So my question for you guys is why we do not get a performance boost when we add the second M2 and why the read speeds are so low?
And it gets down to what component is stopping us from having better performance and I can tell you I've tried with an i7 8700 CPU, even with 32 GB of ram, speeds increase by only a few MB.
What is the bottleneck, is this done on purpose by QNAP? I find it impossible to have such bad performance if you would build a similar machine running windows on the same hardware for example.
Is there a protocol issue?

In attachment several tests with standard config, then with 32 GB RAm and then with 32 GB RAM and i7 8700, as you can see the difference is minimal.

[HiTach]

Watching this thread...

[cooperivan]

My god... I have literally the same setup and literally the same numbers... it was a big disappointment when I got those results...

Any ideas?

[rafale]

Did you test with other hardware to partition and find out what the bottleneck is?
1000MB/s is about right for 10Gb link. Remember that 1MB = 8Mb so a 10Gb link limit will be around 1200MB/s and then you also need to tune your network to make full use of the speed (jumbo frame for example depending of chipset used)
I don't know much about thunderbolt since I pull my card out. I am not seeing the read/write difference you are observing either. I don't think the limiter is your SSD so adding a second drive in Raid0 will do nothing. Even with a PCIex2 link you should be able to hit >1200MB/s in both direction on the drive alone. I am suspicious of your mac, your network and the thunderbolt card.

You can see the results of the read test for my 2 M.2 ssd on mine. They hit 1600MB/s.

Screen Shot 2020-12-04 at 08.45.34.png

[QNAPDanielFL]

If you want to see if that limit is just on the NAS, you can connect to MACs or a Mac through thunderbolt and a PC through 10GbE at the same time. Then see if the combined speed is more than 1000-1300MB/s.

I don't think we would limit speed on purpose. IP over thunderbolt may not be perfect, but those speeds should be great for 4K video editing. I think that is what we were going for when we made Thunderbolt NAS.

1000-1300MB/s does not seem too bad to me for a connection to a single PC. A lot of people have been saying that cache does not increase performance. But 1000-1300 is significantly faster than 4 HDDs RAID 5. So I am actually pleased to see that cache is increasing your performance. But I suspect you might see more throughput from the NAS if you read or write to 2 PCs at the same time.

[HiTach]

@QNAPDanielFL
Do you see the fast write, slow read behavior in-house?
I see the same with access from my Windows 10 desktop/laptop to my 872XT systems over TB3, so switched to 40Gb Ethernet where the speeds are more balanced.
Like the posters above, I just don't understand why read speeds HALF as fast as write speeds.

[QNAPDanielFL]

I was talking with someone about cache performance a while ago. From our conversation.

"quite impossible to benchmark SSD cache accelerated HDD volumes well... most of these benchmark tools write new data each time the test restarts so the cache never has chance to catch up"

To test cache Read speed, you must read the same file 2 times. The first time is HDD speed. Then it copies to the cache and the second time it should read from the cache to test cache speed. Is your AJA test reading the same file more than one time? Maybe your AJA is not able to test cache read speed because it never reads the same file twice.

You can run the test. Set cache to read-only. And then look at your cache hit rate. That should help you know if you are really testing your cache read speed or just HDD read speed.

[rafale]

Sorry if I missed something but I am completely confused by the SSD cache discussion. I don't think the people reporting the problem are using it.
They are talking about direct accessing the SSD.
Cache acceleration as you described is a corner case application. It is completely absurd for most NAS use cases and only causes accelerated SSD and performance degradation. We rarely access the same file more than once when using the NAS as a media center (say you watch a movie) or even as a file sharing (multiple edits from various clients).

The problem people are reporting appear to be related to the transfer rate of the TB card.

[QNAPDanielFL]

If the speed numbers shown with the much faster write performance than read are from an M.2 SSD storage pool, then the reads should be faster than that. I assumed those were the speeds of an HDD storage pool with SSD cache. If it were cache, then the AJA test could be using the cache for writes and not for reads and that could explain the speed test numbers shown.

But if this is an SSD storage pool, it should have faster reads. The fastest read I see in the speed test is less than 600MB/s and it should be over 1000MB/s when connecting to a high-end Mac through thunderbolt. And at least faster than you are getting when connecting to a mid-end MacBook. I don't know what is wrong but those read speeds are not the expected speeds of a thunderbolt connection to our NAS.

[Bob Zelin]

Sorry if I missed something but I am completely confused by the SSD cache discussion. I don't think the people reporting the problem are using it.
They are talking about direct accessing the SSD.
Cache acceleration as you described is a corner case application. It is completely absurd for most NAS use cases and only causes accelerated SSD and performance degradation. We rarely access the same file more than once when using the NAS as a media center (say you watch a movie) or even as a file sharing (multiple edits from various clients).

The problem people are reporting appear to be related to the transfer rate of the TB card.

Rafale is 100% correct here.

I build QNAP systems for professional video editors. The main concern is performance of the QNAP (or any other system) with 4K, 6K and 8K video. I have answered this question over and over and over, but every so often (and here we go again) - someone tries to buy the cheapest box they can, and then complains that they cannot get 800 - 1000 MB/sec performance. Want to get this kind of performance ? ITS EASY - SPEND THE MONEY - BUY THE RIGHT PRODUCT. If you get an 8 bay QNAP like a TVS-872XT or the new TVS-h1288X, or get real fancy and get the larger TVS-h1688X, then you will get EXACTLY the performance that you want. But over and over and over, I see these guys buy a TS-453BT3 (or worse), or the 4 bay TVS-472XT, or even worse - the guy that buys the TVS-872XT, but only buys 3 SATA drives for it - and then he CRIES that he can't get the great performance.

WELL - there is an easy solution - BUY THE RIGHT PRODUCT, and you will have no issues. I don't care if this is QNAP or not. I see this with guys that go out and by a a new iMac Pro, and buy a 2 bay G-Tech Thunderbolt 3 RAID array, and cry "I can't play back my videos" - BUY THE RIGHT PRODUCT, and everything will work. Make the investment. And stop your damn whining.

Bob Zelin

[rafale]

If the speed numbers shown with the much faster write performance than read are from an M.2 SSD storage pool, then the reads should be faster than that. I assumed those were the speeds of an HDD storage pool with SSD cache. If it were cache, then the AJA test could be using the cache for writes and not for reads and that could explain the speed test numbers shown.

But if this is an SSD storage pool, it should have faster reads. The fastest read I see in the speed test is less than 600MB/s and it should be over 1000MB/s when connecting to a high-end Mac through thunderbolt. And at least faster than you are getting when connecting to a mid-end MacBook. I don't know what is wrong but those read speeds are not the expected speeds of a thunderbolt connection to our NAS.

Clearly the OP talks about testing by adding a SSD in a RAID array. He seems to be using neither storage pools nor SSD caching. It is a simple SSD RAID array. You are correct that the read numbers are unexpected which is the reason for this “performance issue” post. I am not seeing this problem but then again, I am not using QNAP NIC or QNAP TB card.

[dwight1379]

There are a number of speed test on YouTube, have you compared your set up with any on there?